Why So Few Firms Can Point to AI-Driven Productivity Gains

Relatively few firms so far have been able to quantify artificial intelligence productivity gains. But that has been the case for computing in general and the internet: it takes time for innovations to transform business processes. In other words, associated and complementary intangible capital also must be created. 

Studying the productivity impact of computerization on 527 large U.S. firms over 1987-1994, professors Erik Brynjolfsson, MIT Sloan School of Management and Lorin Hitt, University of Pennsylvania, Wharton School found that computerization measured over a five-year to seven-year period found productivity and output contributions up to five times greater than over a one-year period. 

“The results suggest that the observed contribution of computerization is accompanied by relatively large and time-consuming investments in complementary inputs, such as organizational capital,” the researchers say.

We could note the same trend with regards to the internet: productivity did not improve quickly, as whole business processes had to be revised. 

Study	Period / data	What it found	Relevance to the paradox
Brynjolfsson and Hitt, 1996, "Paradox Lost?"	Firm-level IT spending	Found substantial returns to information systems investment at the firm level despite weak aggregate evidence.	Early evidence that the "paradox" could be a measurement or aggregation problem. (PubsOnline)
Brynjolfsson and Hitt, 2000, "Beyond Computation"	Firm-level/case evidence	IT's value depended heavily on organizational transformation and intangible investments.	Probably the most important conceptual explanation for why technology's benefits arrive with a lag. (American Economic Association)
Brynjolfsson and Hitt, 2003, "Computing Productivity"	~600 U.S. firms, 1987–94	Returns to computers were 2–5 times greater over seven years than one year.	Strong evidence that complementary investments take years to generate their full payoff. (ResearchGate)
Oliner and Sichel, 2000	U.S. economy	IT accounted for roughly two-thirds of the acceleration in productivity growth between the first and second halves of the 1990s.	By the late 1990s, the productivity payoff of IT had become visible at the macro level. (Federal Reserve)
Stiroh, 2002	61 U.S. industries	Productivity acceleration was greatest in IT-producing and IT-intensive industries.	Shows diffusion beyond the technology-producing sector. (Federal Reserve Bank of New York)
Barua et al., 2004, "Net-Enabled Business Value"	>1,000 firms	Internet-enabled capabilities improved operational performance and ultimately financial performance; supplier/customer readiness mattered greatly.	Direct evidence that Internet adoption plus complementary organizational capabilities generated business value. (AIS eLibrary)
López Sánchez et al., 2006	464 Spanish firms	Both IT investment and workplace Internet use were associated with higher productivity.	Direct firm-level evidence of an Internet-productivity relationship. (ScienceDirect)
Bloom, Sadun and Van Reenen, 2007/2012	U.S. and European multinationals	U.S. firms obtained substantially greater productivity from IT, largely because of superior management practices.	Powerful evidence that management complements technology. (National Bureau of Economic Research)
Quirós Romero and Rodríguez Rodríguez, 2010	2,168 Spanish manufacturing firms, 2000–05	E-buying significantly improved firm efficiency.	Shows that specific Internet-enabled processes, rather than "Internet adoption" generally, mattered. (ScienceDirect)
Huang and Liu, Taiwan e-commerce study, 2013	Taiwanese manufacturing firms, 1999–2002	E-commerce and R&D both raised productivity; their combination was complementary, with network effects.	Internet value increased when combined with other forms of innovation. (ScienceDirect)
Najarzadeh, Rahimzadeh and Reed, 2014	108 countries, 1995–2010	Internet use had a statistically significant positive relationship with labor productivity.	Evidence that the Internet's productivity effects eventually appeared at the macro level. (ScienceDirect)

Beyond all that, some productivity enhancements are difficult to measure, especially when the capabilities do not have a price tag, and are usable without extra charge, such as search, email, navigation, maps or  social media. 

How do we capture the value of increases in consumer choice, reduced transaction costs, reduced search costs, easier price comparison, better product matching or enhanced convenience?

A corollary might be that the Internet and other general-purpose technologies such as electricity become less visible precisely as their economic importance increases. In other words, the technology becomes embedded in all products and services and becomes less visible as a result. 

So organizational performance enhancements are increasingly difficult to isolate from overall organizational prowess. 

In the case of AI, organizations might already be getting substantial value from AI through:

• employees completing tasks faster

• better-quality work

• broader scope of work

• fewer errors

• faster customer responses

• employees handling more work without additional hiring

• better and faster software development

• faster research and analysis

• improved sales and marketing.

But such improvements are tough to quantify; more qualitative than quantitative in terms of output. 

The upshot is that we should not be surprised when few organizations can point to quantitative output gains using accounting practices. First of all, it is too early for the big results to be proven. Also, some of the immediate gains are difficult to impossible to quantify in output, cash flow or profit figures. 

It will take time, even if investors are impatient.

Study Says AI Reshaping Labor Markets

A new Dallas Federal Reserve study suggests generative artificial intelligence adoption is reshaping the Texas labor market, primarily by reducing demand for some types of work, not so much by eliminating existing positions but by reducing demand for recent college graduates. 

“The most exposed occupations are generally in software development, web design and other computer-heavy occupations,” the report says. “Managers, clerical workers, editors and other white-collar occupations are also subject to some of the highest levels of AI task exposure.”

source: Dallas Federal Reserve

“There is strong evidence that GenAI has decreased labor demand for occupations consisting of tasks that can be performed by these new tools,” the researchers say. Still, “the overall effect on aggregate online job posting behavior thus far has been modest.”

“Recent college graduates, whose unemployment rate rose to unusually high levels during this period of rapid GenAI adoption, are where effects of GenAI on employment and earnings are likely to first appear,” the study suggests. 

So far, AI's labor-market effect is that “fewer people get hired into certain jobs,” rather than “large numbers of existing workers fired from those jobs.” 

Study	Data / period	Main finding	Job impact
Brynjolfsson, Chandar & Chen — “Canaries in the Coal Mine” (Stanford, revised 2026)	ADP payroll data, millions of workers, through June 2026	No widespread economy-wide displacement, but employment of 22–25-year-olds in highly AI-exposed occupations is 19% below the counterfactual. Experienced workers show no comparable gap.	Strong evidence for the hiring channel. The authors say the adjustment occurs primarily through reduced hiring of young workers rather than increased separations. (Stanford Digital Economy Lab)
Hosseini Maasoum & Lichtinger — “Generative AI as Seniority-Biased Technological Change” (Harvard, 2026 revision)	65 million résumés, 280,000+ firms	Junior employment falls following GenAI adoption, particularly in highly exposed occupations; senior employment is largely unchanged.	Decline is driven primarily by slower hiring rather than increased separations. (SSRN)
Tucker — “You're (not) Hired” (U.S. Census, 2026)	Matched employer-employee administrative data	Early-career employment in the most AI-exposed industry/state cells fell 12% over 10 quarters after ChatGPT.	The paper finds the decline in employment was primarily caused by a large decrease in hiring. This is perhaps the clearest administrative-data evidence of the mechanism. (Census.gov)
Liu, Wang & Yu — “Labor Demand in the Shadow of Generative AI” (World Bank, 2026 revision)	285 million U.S. online job postings, 2018–2025	Postings for occupations highly vulnerable to AI substitution fell 9% relative to less-vulnerable occupations, with the differential reaching 15% by the third year.	Direct evidence of reduced labor demand, rather than layoffs. Particularly important because it examines vacancies at enormous scale. (SSRN)
Audoly, Guerin & Topa — New York Fed, “Do Job Postings Show Early Labor-Market Effects of AI?” (2026)	U.S. Lightcast postings	Overall hiring has slowed, but they find little evidence of a distinct AI-driven decline in postings for AI-exposed occupations.	Important counterweight: the aggregate slowdown in postings cannot confidently be attributed to AI. (Liberty Street Economics)
Federal Reserve Board — “AI Adoption and Firms' Job-Posting Behavior” (2026)	Firm/industry AI adoption + job postings	No evidence that firms or industries with greater AI adoption have reduced total job postings.	Suggests that if AI is eliminating some positions, firms may be switching hiring toward other jobs, rather than simply reducing total hiring. (Federal Reserve)
Gimbel, Kendall & Nunn — Yale Budget Lab (2026)	Monthly CPS employment/wage data	After controlling for differences between exposed and unexposed occupations, they find no statistically or economically significant aggregate employment or wage effect.	Again, little evidence of broad existing-job destruction. The effects may be concentrated in particular populations. (The Budget Lab)
Humlum & Vestergaard — “Still Waters, Rapid Currents” (NBER, 2025/26)	Danish administrative records + AI adoption surveys	No detectable effect on earnings or hours, even among early adopters and highly exposed workers. But substantial task restructuring and occupational switching occurred.	Little evidence of job elimination so far. Firms appear initially to be reorganizing work rather than cutting employment. (National Bureau of Economic Research)
Chandar — “Tracking Employment Changes in AI-Exposed Jobs” (2025)	U.S. CPS, Q4 2022–Q1 2025	No substantial aggregate employment/earnings difference in highly exposed occupations, although software and customer-service occupations diverge.	Overall employment effects small; evidence of localized employment declines, not economy-wide displacement. (SSRN)
Frank et al. / related job-posting research summarized by Stanford	U.S. job postings	Several studies find greater declines in postings in AI-exposed occupations. But some declines began before ChatGPT and correlate with interest rates/remote work.	Supports reduced postings, but attribution to GenAI is contested. (Brookings)
“Winners and losers of generative AI” (JEBO, 2025)	Online freelance marketplace	About 10% of postings were judged substitutable by GenAI; demand for those skill clusters fell as much as 50% in short-term roles.	Strong evidence of demand substitution in particular tasks, although aggregate freelance demand did not fall. (ScienceDirect)
PwC 2025 AI Jobs Barometer	Lightcast job postings, 2019–2024	U.S. occupations with greater GenAI exposure experienced substantially slower job-posting growth: roughly 2% vs. 20% for less-exposed occupations.	Strong descriptive evidence of slower demand growth, though not necessarily causal evidence of AI. (PwC)

But trends could, or maybe, should, change over time, as actual job displacement or elimination, plus creation of new jobs, happens. 

Effect	Evidence so far
Mass layoffs caused by GenAI	🔴 Little evidence
Economy-wide employment decline	🔴 Little/no evidence
Reduced total hiring because of AI	🟡 Weak/mixed
Reduced hiring in particular AI-exposed occupations	🟢 Increasing evidence
Reduced entry-level hiring	🟢 Significant evidence, but causality disputed
Reduced job postings in AI-substitutable occupations	🟢 Fairly strong evidence
Task substitution/reorganization	🟢 Strong evidence
Productivity increases without employment reductions	🟢 Strong evidence
Wage effects	🟡 Generally small so far
Long-term displacement	❓ Still largely unknown

On balance, it is possible net job creation could happen, as AI creates new jobs and roles, despite AI-induced reductions. 

Study	Period / data	Main finding	Implication
Autor, Chin, Salomons & Seegmiller, “New Frontiers: The Origins and Content of New Work” (QJE, 2024)	U.S., 1940–2018; ~35,000 Census occupations	About 60% of employment in 2018 was in occupations that did not exist in 1940. New work emerged alongside technological change, particularly in professional and service occupations after 1980. (National Bureau of Economic Research)	Perhaps the strongest evidence that technology doesn't simply redistribute a fixed number of jobs. It creates new categories of work.
Autor, “Why Are There Still So Many Jobs?” (JEP, 2015)	Historical review	Automation substitutes for labor in particular tasks but also complements workers, raises productivity and output, and increases demand for labor elsewhere. (TopCat)	Explains why technological progress can eliminate particular jobs without eliminating work generally.
Acemoglu & Restrepo, “Automation and New Tasks” (JEP, 2019)	U.S. historical/employment evidence	Automation creates a displacement effect, but creation of new tasks produces a reinstatement effect that raises labor demand. (AEA Publications)	Provides a theoretical framework for understanding why job creation can offset automation.
Acemoglu & Restrepo, “The Race Between Man and Machine” (AER, 2018)	Long-run economic model	Technology that automates existing tasks reduces labor demand, while creation of new tasks has the opposite effect. (American Economic Association)	Net employment depends on the balance between automation and new-task creation.
Bessen, “Automation and Jobs: When Technology Boosts Employment” (2017)	Historical industries, including textiles, steel and automobiles	Industries experiencing rapid productivity growth sometimes experienced employment growth, because lower costs stimulated demand sufficiently to offset labor-saving technology. The OECD reviews this evidence. (OECD)	Lower prices can create enough additional demand to more than compensate for labor-saving technology.
Mann & Puttmann, “Benign Effects of Automation” (2018)	U.S. counties/industries, patent data	Automation innovations were associated with declining manufacturing employment but increasing employment in services. (OECD)	Job destruction and creation can occur in completely different industries and locations.
Autor & Salomons, “Is Automation Labor-Displacing? Productivity Growth, Employment, and the Labor Share” (2018)	28 industries, 19 advanced economies, 1970–2015	Productivity improvements can reduce employment within an industry, but positive spillovers to other industries more than offset those losses. The OECD summarizes the evidence. (OECD)	This is especially important: you shouldn't look only at the industry where automation occurs.
OECD, Employment Outlook 2019	OECD countries, long historical perspective	Despite substantial technological displacement, overall employment has generally grown. The OECD concludes that historically the net effects of major technological revolutions on employment have been positive. (OECD)	Broad institutional review supporting the historical pattern.
OECD, Technology, Productivity and Job Creation (1998)	OECD historical evidence	Technological change destroys jobs in some industries while creating jobs in others; historically the process produced net job creation as new industries replaced old ones and demand expanded. (OECD)	An earlier, broad cross-country examination reaching the same conclusion.